Licensing Company Data for AI Training Under the GDPR (2026)
35 min readAIData.Broker research18 Oct 2026132 records, 82 sources
Yes. A company in the EU, the UK or Switzerland can license its own business records to an AI buyer, including one in the United States, and no law forbids it. But every record that names or can identify a living person is personal data, and then four things must be true before the licence is signed. The holder needs a lawful basis, in practice legitimate interests, backed by a written compatibility test for the new purpose; the people in the records must be told in advance and be able to object; sensitive data such as health or union membership must be kept out; and identifiable data sent to a US buyer needs a transfer tool. The closest ruling, from the EU's Court of Justice, accepts that a commercial interest can be a legitimate interest when an organisation discloses members' data for money, but asks whether the disclosure is strictly necessary and what people could expect. That is why the cleanest route is to de-identify first, so that what leaves the holder is anonymous, or close to it, in the buyer's hands. Records about companies rather than people are outside the GDPR altogether.
This guide is for the founder, managing director or data protection officer of a company in the EU, the UK or Switzerland deciding whether to license its records (support tickets, chat logs, email, documents, code) to an AI buyer. It covers the EU's General Data Protection Regulation (the GDPR), the UK GDPR as changed by the Data (Use and Access) Act 2025, and Switzerland's Federal Act on Data Protection (the FADP), and ends with two worked examples, a checklist and the questions people ask. For buyers, tax and payment, read selling data to US AI companies from outside the US.
This guide is research from public sources, not legal advice.
Contents
- The answer in one table
- Sort your records first
- Lawful basis and the compatibility test
- Special categories and staff records
- Telling people and the right to object
- Anonymisation and pseudonymisation
- Sending records to a US buyer
- What regulators and courts have done on AI training
- The rights that give a licence its value
- The Data Act against one-sided buyer terms
- The AI Act: the training data summary
- Worked examples: a German software company and a Dutch online shop
- The checklist
- Questions people ask
- What this means for you
1. The answer in one table
The three regimes share one structure: the UK GDPR began as a copy of the EU text, and Switzerland revised its Act in 2023 to sit close to it. Each asks whether data is personal, whether it may serve a new purpose, whether people were told and whether it may leave the country. Three differences matter most. The UK's 2025 Act rewrote the purpose test, the research rules and the transfer test, and those changes have applied since 5 February 2026. Switzerland has no free-standing legitimate interests basis: processing that breaches a person's personality rights is unlawful unless justified by consent, an overriding interest or law. And the fines differ by an order of magnitude.
| Question | EU (GDPR) | UK (UK GDPR after the 2025 Act) | Switzerland (FADP) |
|---|---|---|---|
| Is data only about companies covered? | No: legal persons are outside it | No, the same definition | No: natural persons only since 2023 |
| Usual basis for a licence | Legitimate interests, a three-part test | Legitimate interests; the new recognised legitimate interests do not cover licensing | An overriding private interest, consent or law |
| Test for a new purpose | Article 6(4), five factors | Article 8A, five factors | Compatible with a purpose the person can recognise |
| Telling people | Before the new use starts | Before, with a narrow research exception | Check the Act's information duty (not covered here) |
| Sending to a US buyer | A buyer on the Data Privacy Framework list, or standard clauses | A buyer in the UK Extension, or the IDTA or the Addendum | A buyer certified under the Swiss-US framework, or approved clauses |
| Fine ceiling | 20 million euro or 4 percent of worldwide turnover | 17.5 million pounds or 4 percent of worldwide turnover | 250,000 francs for a wilful unlawful disclosure abroad, on the person responsible |
The de-identify-first route
Every condition in the table gets easier when the people are taken out of the records before anything leaves. Legitimate interests fails, in the UK regulator's words, if you can reasonably achieve the same result in another less intrusive way, and a buyer training a model on support conversations needs the wording, not the customers' names. Even research must use data that no longer identifies people where possible, and truly anonymous data is outside the Regulation, so no transfer tool is needed for it. The law does not disappear for the holder, though: it keeps the originals, the anonymising step is itself a further processing, and a claim of anonymity must be shown, not asserted.
2. Sort your records first
The GDPR applies to personal data: any information relating to an identified or identifiable natural person. A ticket from a named customer, a chat with an email address and a log with an IP address are personal data; a price list, product documentation and code that holds no personal data are not. The Regulation says in terms that it does not cover data about legal persons, including the name and the form of the legal person and the contact details of the legal person, and Switzerland has protected natural persons only since 2023.
That gives a holder a first, cheap win: much of what buyers value (how the product works, how problems were solved) can be separated from the people in it. But the person who opened a ticket and the engineer who closed it are natural persons, and their details are personal data even when they wrote at work.
The second sort is about whose records they are. Only the controller, the business that decides why and how data is processed, can decide to license it. A software company that holds its customers' data inside its product is their processor, and a processor acts only on documented instructions from the controller. If it licenses that data anyway, it shall be considered to be a controller for that processing and has infringed the Regulation. The line runs through most software companies: tickets sent to its own support desk are its own records, while the files and end-user data customers keep inside its product are theirs. Buyers' counsel accept customer content only if it is excluded, expressly permitted by each customer contract, or anonymised.
| What the records are | Whose they are | Which rules apply | What to do |
|---|---|---|---|
| Product documentation, code, price lists, records about companies | Yours | Contract, copyright and trade secrets, not data protection | License them; check confidentiality owed to others |
| Your own support tickets, sales email, customer chat | Yours, as controller | The GDPR in full | De-identify, assess, tell people |
| Data you process for customers | Your customers', with you as processor | Article 28 | Leave out unless each customer authorises it |
| Staff email, chat, HR files | Yours, about your staff | The GDPR plus national employment rules | Leave out, or de-identify and involve staff representatives |
| Documents others wrote, partner material | Often someone else's copyright or secret | Copyright and contract | Leave out or clear |
3. Lawful basis and the compatibility test
Every use of personal data needs a lawful basis, and a licence is a new use. For records already held, consent is rarely practical: customers from eight years ago cannot all be reached, and silence is not agreement. For staff, the European regulators consider consent unlikely to be freely given at all. That leaves legitimate interests, which allows processing necessary for the purposes of the legitimate interests pursued where the people's rights do not override them.
It is a test, not a label. The European Data Protection Board (the EDPB, the EU regulators' joint body) expects controllers to have carefully assessed and documented three cumulative conditions: a legitimate interest, necessity, and a balance that does not tip against the people. If the balance fails even after safeguards, the basis is simply unavailable. France's regulator accepts a commercial interest for AI development on the same three conditions, and a German appeal court accepted that economic interests can be legitimate for AI training.
The nearest case: a federation that disclosed members' data for money
The closest ruling to a holder licensing records is the Court of Justice's October 2024 judgment on the Dutch tennis federation (KNLTB), which had passed members' data to sponsors for payment. A commercial interest could constitute a legitimate interest, the Court held, but the federation could have informed its members beforehand and asked them, a less intrusive means, and what members could reasonably expect at collection carries particular importance. For an archive collected under notices that said nothing about licensing, this points toward de-identifying or asking first.
The compatibility test
Separately, personal data may not be further processed in a manner that is incompatible with the purpose it was collected for. Unless licensing was announced at collection, the holder runs the five-factor test in Article 6(4): the link between the purposes, the context and relationship, the nature of the data, the possible consequences, and safeguards such as pseudonymisation.
The GDPR's recital 50 says a compatible purpose needs no legal basis separate from the original one. Two authorities read it more strictly. The French regulator (the CNIL) says the test covers sharing data with third parties and that a compatible purpose still needs a valid legal basis and information to people; the Higher Regional Court of Cologne (OLG Köln) says Article 6(4) gives no permission of its own. Do both, in writing, dated and per lot. The CNIL also says that where a third party supplies data collected for another purpose, it is for the third party to ensure the transmission is compatible, and a licence can limit reuse to keep it so. The holder is that third party, so a licence naming AI model training as the only purpose, with no re-identification or onward sale, is part of its own case.
Expectations are judged at collection: AI wording added to a privacy policy later does not necessarily mean people can expect it, and the Cologne court, refusing an injunction against Meta's training in May 2025, found no such expectation for data posted before the training was announced. A holder's oldest records are where the balance is hardest to win.
The UK after the Data (Use and Access) Act 2025
The Act's new fast-track basis, recognised legitimate interests, covers public-task requests, national security, emergencies, crime and safeguarding, and its examples of ordinary legitimate interests do not include licensing, so a UK holder stays on the three-part test. The Act replaced Article 6(4) with Article 8A, keeping five factors, and lists purposes that count as compatible automatically ( consent, research under Article 84B and Annex 2); commercial licensing is not among them. The regulator (the ICO) says a very different or unexpected purpose means you're likely to need to ask for specific consent. One ICO position helps holders: it expects AI developers to evidence why other available methods for data collection are not suitable before relying on scraping, and a documented licence is that alternative.
Switzerland
The Swiss Act allows collection only for a specific purpose that the data subject can recognise, with further processing compatible with it. With no separate legitimate interests basis, a holder relies on an overriding private interest, and disclosing sensitive data to third parties is a listed breach. De-identification makes the overriding-interest argument much easier.
Research: an argument, not a basis
The GDPR treats research reuse as not incompatible with the original purpose, its recital 159 reads research broadly, and the UK now includes commercial research, with safeguards in Article 84B. But no regulator has said licensing to a commercial buyer is research, and the CNIL says training a model is not a statistical purpose just because it uses statistics. Treat research as an argument a buyer may make for its own processing, not as the basis for your licence.
4. Special categories and staff records
Some data is prohibited unless a narrow exception applies: data revealing racial or ethnic origin, political opinions, religious or philosophical beliefs, or trade union membership, genetic data, biometric data used to identify someone, and data about health, sex life or sexual orientation. Tickets and chats contain these more often than holders expect: a customer explaining a late payment by illness, an employee mentioning a union meeting, a delivery moved for a religious holiday. None of the exceptions is written for a commercial licence.
The Court of Justice made this a whole-corpus problem. Where a set mixes sensitive and ordinary items collected en bloc, processing the set must be regarded as being prohibited unless an exception applies. A holder cannot rely on the sensitive share being small: it must screen content, have a person read what the screen flags, and remove those items before the lot is built. Call recordings are the hardest case, the records least likely to de-identify cleanly, and most holders leave them out.
Staff data carries a second layer. Each member state may adopt more specific rules for it, and national provisions that do not meet the GDPR's conditions must be disregarded, so a national rule or works agreement does not automatically make a use lawful. Staff representatives can also have a say. In Germany a works council co-determines the introduction and use of technical systems designed to monitor staff conduct or performance, and an export showing what each employee did, or the tooling that processes it, can come within that right. In the Netherlands, buyers' counsel ask whether the works council approved the processing arrangement. The safest course is to leave staff email, chat and recordings out, or de-identify them, and replace agents' names in tickets; the guide to employee emails and chats goes further.
5. Telling people and the right to object
A holder relying on legitimate interests must tell people and let them say no. For its own customers and staff, it must provide information about a new purpose prior to that further processing; reusing first and informing later is not allowed. For people it never dealt with (a third party named in an email), the duty yields where informing proves impossible or would involve a disproportionate effort, but Italy's regulator wants a specific, documented assessment and public information before that exception is used.
Anyone whose data is processed on legitimate interests may object at any time, and the holder must stop unless it shows compelling overriding grounds. The EDPB lists two fitting safeguards: a reasonable period of time between announcing a use and starting it, and an opt-out that needs no reasons. When Meta resumed EU training in May 2025, Ireland's regulator listed the changes it had made, including an easier objection form, de-identification and filtering.
In practice a holder needs, before the lot is built: an updated privacy notice; a direct notice to every customer and member of staff it can reach; a public notice for the rest, with the disproportionate-effort reasoning on file; an objection form that needs no reasons; a waiting period; and a way to pull one person's records out of the lot. Objections must be honoured before delivery, because once a lot is anonymous in the buyer's hands nobody can find one person in it. An impact assessment (the written risk review for high-risk processing) is also likely, since processing with new technologies likely to result in a high risk needs one. The UK excuses the duty to inform only for research under Article 84B where informing is impossible or would involve disproportionate effort, a hard case to make for customers whose email addresses the holder has.
6. Anonymisation and pseudonymisation
Three states of data matter. Identifiable data names people or traces to them easily. Pseudonymised or de-identified data has its direct identifiers replaced while someone, usually the holder, keeps the originals. Anonymous data cannot identify anyone by any means reasonably likely to be used. The GDPR's principles should not apply to anonymous information, while the EDPB says pseudonymised data is to be considered information on an identifiable natural person, even when the data and the key are in different hands.
The Court of Justice has since taken a more relative view. In September 2025, in EDPS v SRB, it held that pseudonymised data must not be regarded as constituting, in all cases and for every person, personal data: for a recipient that cannot re-identify anyone, it may not be. The UK regulator agrees that the same information can be personal data in your hands but anonymous in the hands of other parties. Regulators may still apply the stricter view, so write the contract as if the lot were personal data.
What it takes to be anonymous
An anonymisation must prevent all parties from singling out an individual, from linking records about them and from inferring information about them. Free text identifies by context: "our only office manager in Rotterdam", "back from her surgery", a rare product configuration. The UK regulator suggests a "motivated intruder" test, asking what a determined person with reasonable means could work out, and the EDPB says a mere assertion of anonymity is not enough. The anonymising step needs a lawful basis and a stated purpose of its own; it usually passes, but record it and mention it in the notice.
De-identification that falls short of anonymity still counts. The Cologne court treated Meta's removal of names, emails, phone numbers, ID numbers and addresses as a real reduction of risk while agreeing it was not anonymisation, and the EDPB lists masking personal data or substituting it with fake personal data as suited to language model training, where the wording matters and the identities do not.
| State of the lot | For the holder | For the buyer | Transfer tool to a US buyer |
|---|---|---|---|
| Identifiable | The GDPR in full | The GDPR in full, as a new controller | Yes |
| De-identified or pseudonymised | Still personal data (it keeps the originals) | May be non-personal if it cannot re-identify; regulators may say otherwise | Treat it as needed |
| Anonymous | The anonymising step is processing; the result is outside the law | Outside the law | No |
Who de-identifies, and where
Many buyers offer to do the work, and where they do it matters. Mercor anonymises with its own pipeline and micro1 removes personal information itself, after receiving the data, while Scale anonymises inside your own systems before extraction. If identifiable records reach a US buyer first, every condition in this guide applies to the identifiable version; de-identifying before anything leaves keeps the lot on the easier side of the table. Our de-identification standard calls data anonymous only for a named recipient after an independent test, and the de-identification tool runs on your own computer, so the originals never leave it.
7. Sending records to a US buyer
The GDPR's transfer rules apply when there is an exporter subject to the GDPR, a disclosure of personal data, and an importer in a third country. Sending a truly anonymous lot is not a transfer. A lot that is only de-identified is the open case: EDPS v SRB judged identifiability from the recipient's side, but the Court has not said how that applies to transfers, so put a tool in place anyway. Buyers know this: Turing flags early any jurisdiction with data-export restrictions, and Ooak keeps EU partners' data processed and stored only in the European Union.
For EU data there are two practical routes. The first is the EU-US Data Privacy Framework (the DPF): the Commission's 2023 decision covers only transfers to US organisations included in the Data Privacy Framework List, so check the buyer's name and active status on the official DPF list before sending. The second, for a buyer not on the list, is the standard contractual clauses (the SCCs), the Commission's model contract, which supply the appropriate safeguards a transfer without adequacy needs. A licence is a controller-to-controller transfer, Module One of the clauses, which binds the buyer to process the data only for the specific purpose(s) of the transfer listed in its Annex I.B. Name AI model training there and in the licence; the buyer cannot widen it later without the people's consent. The Article 49 fallbacks are interpreted restrictively, and the one resting on the exporter's interests needs transfers not repetitive and concerning only a limited number of data subjects; do not plan on them.
UK data can go to a US buyer without further tools only if the buyer is on the DPF list as participating in the UK Extension. Otherwise the holder uses the UK's International Data Transfer Agreement (the IDTA) or the UK Addendum to the EU clauses, with a transfer risk assessment against the new test that protection abroad is not materially lower. Switzerland allows disclosure abroad with an adequacy decision or approved safeguards, and since 15 September 2024 US companies certified under the Swiss-US framework receive Swiss data on an adequate footing. Each certification is separate, so check the one for your country.
| Holder based in | Buyer certified for your country | Buyer not certified |
|---|---|---|
| EU | DPF list: no further tool | SCCs, Module One, with AI model training named |
| UK | DPF list with the UK Extension | IDTA, or the UK Addendum to the SCCs, plus a transfer risk assessment |
| Switzerland | Swiss-US framework certification | Clauses approved for Switzerland |
If the Data Privacy Framework falls
The DPF survived its first challenge: the EU's General Court dismissed the action in its entirety in September 2025, in a case brought by Philippe Latombe. He has asked the Court of Justice to annul the decision, and the appeal is pending. This has happened before: in 2020 the Court held that the DPF's predecessor, the Privacy Shield, is invalid, while leaving the standard clauses in place. A holder relying on a buyer's listing should write the SCCs into the licence as a fallback that applies if the buyer leaves the list or the decision falls, so deliveries do not stop overnight. What such a ruling would mean for the UK bridge, a separate UK regulation built on the same US list, is not settled.
8. What regulators and courts have done on AI training
Regulators have so far acted on AI developers, not on holders that licensed records. Their decisions still show what a holder will be asked, because whoever deploys a model should assess the source of the data, and the buyer passes that question on.
| Who | When | What happened | What it means for a holder |
|---|---|---|---|
| Ireland's regulator and Meta | June 2024 | Meta paused EU training on public posts | Regulators act before training starts |
| Hamburg's regulator | July 2024 | Training must comply with data protection regulations, but a defect does not taint later use of the model | A softer view the EDPB does not share |
| The EDPB | 17 December 2024 | Opinion 28/2024: legitimate interests can support AI development if documented | Buyers will ask for the assessment, notices and de-identification report |
| Ireland's regulator and Meta | 21 May 2025 | Training allowed after Meta updated its legitimate interest, impact and compatibility assessments | The regulator's list of papers is a holder's checklist |
| Cologne court | 23 May 2025 | Refused an injunction against Meta's training | An interim ruling on public posts |
| Ireland's regulator and LinkedIn | 7 November 2025 | The regulator has not approved, or found compliant the training, but did not intervene further | No regulator has approved any training set |
| Italy's regulator (the Garante) and OpenAI | Removed 2026 | Its 15 million euro decision came down after a Rome court upheld OpenAI's opposition | Fines here are not settled |
| The Garante and Character.AI | 3 July 2026 | Ordered a US developer to find a compatible purpose or delete the data, with a 158,000 euro fine | Check a lot for personal data before handing it over |
The honest gap
No court or regulator has ruled on a holder licensing business records to an AI buyer. The closest decisions are KNLTB, about disclosing members' data for a fee, and the Meta and LinkedIn statements, about developers training on their own users' data, and we found no published enforcement against a holder. That is a gap in the law, not a clean bill of health: every principle applied to developers (a documented basis, expectations at collection, notice, objection, de-identification) applies to a holder's disclosure as much.
The ceilings are real even where fines so far have been modest. A licence that breaches the principles, people's rights or the transfer rules sits in the GDPR's top tier, up to 20 million euro or 4 percent of worldwide turnover, and corrective powers reach erasing part of the dataset and, in the extreme, the model. That risk lands on the buyer, which is why buyers ask holders to prove a lot was lawfully assembled.
9. The rights that give a licence its value
Data protection decides whether a holder may license. Three other bodies of law decide whether there is anything a buyer has to pay for.
Database right. The EU gives the maker of a database a right against extraction where there has been a substantial investment in either the obtaining, verification or presentation of the contents, and the UK kept it on the same test. Its limit catches many holders: the Court of Justice held in the British Horseracing Board case that "obtaining" means the resources used to seek out existing independent materials, not those used to create them. Tickets and chats are created by running the business, so a raw export may carry no database right, while a curated, cleaned corpus is a stronger claim.
Trade secrets. Private tickets, code and logs can be trade secrets if they are secret, valuable because they are secret, and protected by reasonable steps. A buyer that uses a secret in breach of a confidentiality agreement or a duty to limit its use acts unlawfully, and in the UK use that is a breach of confidence is unlawful too. Both depend on the licence limiting purpose and barring disclosure; a licence with no confidentiality terms weakens the very status that makes the records worth licensing.
Why the text and data mining exception does not reach private records. The EU exception lets anyone copy lawfully accessible works for mining unless the rightholder opted out in machine-readable form, and the AI Act requires model providers to have a policy to comply with those opt-outs. German courts have applied it to public material both ways: building the LAION image dataset was covered, while a model's memorised song lyrics were not. The UK exception covers only analysis for the sole purpose of research for a non-commercial purpose. None of this reaches a holder's private records: the exception needs lawful access, which a buyer does not have to internal tickets and code. The only lawful route from a holder's systems to a training set is a licence, which is why private records are worth paying for; the guide to scraping covers public data.
10. The Data Act against one-sided buyer terms
For contracts concluded after 12 September 2025, a term on access to and the use of data, or liability and remedies for breaching data obligations, unilaterally imposed by one enterprise on another, does not bind the other if it is unfair. It is not limited to device data: it reaches any private sector data accessed and used on the basis of contract between enterprises, including a licence of company records.
A term is always unfair if it lets the imposing party exclude or limit its liability for intentional acts or gross negligence. A term is presumed unfair if it inappropriately limits remedies or liability, or lets the imposing party access and use the data of the other contracting party in a way significantly detrimental to it. An unfair term falls away, the remaining terms stay binding and the parties cannot contract out.
The limits matter. A term counts as imposed only where the other party has not been able to influence its content despite an attempt to negotiate it, and the rule does not test the price. It is an EU rule, so UK and Swiss holders do not have it. And no court has decided whether it reaches a licence governed by Delaware or New York law, the usual choice in US buyers' forms, though an EU court may apply overriding mandatory provisions whatever law the contract names. Facing a take-it-or-leave-it form, an EU holder should propose changes to the data terms in writing and keep the record, watching for use beyond the licensed purpose, caps on the buyer's liability for misuse or re-identification, and one-sided indemnities. The licence terms register shows what buyers propose, and what is a data licensing agreement? walks through each clause.
11. The AI Act: the training data summary
A provider of a general-purpose AI model must draw up and make publicly available a sufficiently detailed summary of its training content, on a template set by the EU's AI Office, a duty that has applied from 2 August 2025. It is meant to be generally comprehensive in its scope instead of technically detailed.
The template decides whether a holder is named. It does not require specific data to be identified, though providers may voluntarily disclose more. For licensed data it asks only whether such licences exist and their kinds, and private datasets from third parties are named only if already public, otherwise described by kind, personal-data content and language. Separately, the provider keeps documentation for the AI Office that records the type and provenance of data.
So a lot will normally appear in public only as a general description. Say in the licence whether you may be named, because naming is the buyer's choice, and expect the buyer to ask for your rights file; the guide to training data provenance explains what goes in it.
12. Worked examples: a German software company and a Dutch online shop
A 60-person German software company with eight years of support tickets
Example (illustrative, not a real company). A Munich company sells scheduling software to mid-sized businesses. It has 60 staff, a works council, and eight years of tickets in Zendesk, written in German and English by its customers' employees. A buyer has asked to license them.
Sorting. The tickets are personal data, because each names the customer employee who wrote it; only the company's own name, form and contact details fall outside the GDPR. They are the company's own records as controller. Their attachments are not: screenshots and exports of customers' own scheduling data, full of their staff's shifts and absences, which the company holds as processor. Those are excluded. Product documentation and release notes go into the lot on a trade-secret and contract basis.
Basis and staff. The legitimate interest assessment finds a commercial interest, which KNLTB accepts, but necessity fails for identifiable tickets: a model learning from support conversations does not need names. Tickets from before the privacy notice mentioned AI carry no expectation, so de-identification must be strongest there, and the company records both the assessment and the compatibility test. A screen removes tickets where customers' staff explained absences by illness. Agents' names are replaced, and the works council is consulted before the export runs, since the export shows what each agent did and can come within its co-determination right.
Notice and transfer. The company updates its notice, emails every customer contact it can reach, publishes a notice for the rest, opens an objection form, waits several weeks and removes objectors. It replaces names, emails, phone numbers, ticket IDs and customer company names (confidential under its customer contracts), and has the result tested by a provider it chose. The buyer offered to de-identify on its own systems, which would have sent identifiable tickets to the US, so the company delivers only the de-identified lot, with Module One clauses as a fallback. The buyer's form let it combine the data with any source and capped its liability for re-identification, so the company asks for changes in writing, relying on the Data Act's presumptions if refused.
Value. micro1 wants English records, so the German half narrows its buyers, while The General Data Company courts German business leaders and shows prices in euros. Buyers' own estimators price a company in Europe at about 0.71 times of the same company in the US ( the location discount). The Germany page brings the country's rules, tax and buyers together.
A Dutch online shop with customer chat logs
Example (illustrative, not a real company). A Utrecht company sells home and garden goods to consumers. It has six years of customer chats in Intercom, in Dutch and English, and has been asked about licensing them.
The consumer problem. Every chat is personal data about a consumer who had no reason to expect a licence. KNLTB, which began as a dispute with the Dutch data protection authority, is directly on point: expectations at collection carry particular importance, and asking first is the less intrusive means. Asking years of former customers is impractical, so the realistic route is a lot that is anonymous in the buyer's hands.
What comes out. Names, addresses, phone numbers, order numbers and payment references are replaced, and photos of damaged goods, which show homes and sometimes faces, are excluded. A screen flags chats mentioning a disability, a pregnancy or a medical reason for a delivery arrangement; under the mixed-dataset rule the shop cannot rely on such chats being rare, so a person reads and removes them.
Notice, assessment and value. The shop emails its customer list, updates its notice, offers an objection form and waits. Because the lot is a large set of conversations, it runs an impact assessment under Article 35. If the lot passes the anonymity test, sending it is not a transfer, but the licence still carries Module One clauses in case a regulator takes the cautious view. Dutch-language chats reach fewer buyers than English ones, which a valuation should reflect; the Netherlands page has the country's details.
13. The checklist
The first four steps shrink the problem, the next four meet the duties that remain, and the last four deal with the buyer.
| Step | What to do | The rule behind it |
|---|---|---|
| 1. Sort the records | Mark what is yours, what you hold for customers, staff data, company-only data and third-party content | Article 28(10), recital 14 |
| 2. Remove sensitive items | Screen for health, union, beliefs, biometrics; a person reads what is flagged | Article 9(1), the mixed-dataset rule |
| 3. Leave out or de-identify staff content | Check national rules; involve the works council where there is one | Article 88, BetrVG 87(1) no. 6 |
| 4. De-identify and test | Replace names, emails, phones, IDs, card numbers, addresses; test with a motivated intruder; keep the report | EDPB on masking, singling out, linking, inference |
| 5. Write the assessments | A legitimate interest assessment and a compatibility test, dated, per lot | EDPB Opinion 28/2024, UK Article 8A |
| 6. Run an impact assessment | Where the lot is large or holds conversations or recordings | Article 35(1) |
| 7. Tell people first | Update the notice, write to those you can reach, publish for the rest, wait | Article 13(3), Article 14(5)(b) |
| 8. Honour objections | An opt-out with no reasons; remove objectors before delivery | Article 21(1), EDPB on opt-outs |
| 9. Pick the transfer tool | A DPF listing for your country, or SCCs, the IDTA or the Addendum | Article 46(1), the UK bridge |
| 10. Write the licence | AI model training as the only purpose; no re-identification or onward sale; deletion, audit and confidentiality | SCC Clause 8.1, CNIL on supplied data, trade secrets |
| 11. Check the buyer's form | Push back on one-sided data and liability terms in writing; decide whether you may be named | Data Act Article 13, the summary template |
| 12. Keep the file | Assessments, notices, objections handled, de-identification report, transfer tool, licence | the next party checks the source |
The readiness tool asks the same rights questions in order, and the datasheet tool describes the finished lot in the terms buyers use.
Questions people ask
Is it legal under the GDPR to sell data for AI training? Yes, if the four conditions hold: a documented lawful basis (usually legitimate interests), a compatibility test, people told in advance with a right to object, and a transfer tool if identifiable data leaves. Most holders de-identify first, because identifiable records collected under old notices rarely pass the balance.
Do I need my customers' consent? Not necessarily: legitimate interests can work without it if the assessment, notice and objection route are in place. But the UK regulator says an unexpected purpose means you're likely to need to ask for specific consent, and KNLTB named asking first as the less intrusive means. Without consent, de-identification is what makes the balance work.
Is anonymised data covered by the GDPR? Truly anonymous data is outside it. Pseudonymised data is personal for whoever can re-identify it, and since EDPS v SRB it need not be personal for every person, though regulators may take a stricter view. A claim of anonymity must be shown, not asserted.
Can I send personal data to a US AI company? Yes, if the buyer is on the Data Privacy Framework list for your country, or under the standard clauses, which bind the buyer to the purposes you name. Keep the clauses in the licence as a fallback while the appeal against the DPF is pending.
Did the UK's Data (Use and Access) Act 2025 make AI training easier? Not for licensing: its recognised legitimate interests do not cover commercial licensing, and nor does its list of compatible purposes. It widened research to include commercial research, but whether training a commercial model qualifies is untested.
Can a software company license data it holds for its customers? Not on its own decision. A processor acts only on documented instructions from the controller, and one that licenses customer data anyway is treated as a controller and infringes the Regulation. Only each customer can authorise that use.
Will the AI company name my company in its training summary? Usually not. The AI Office template asks only whether licences exist and their kinds. Write your preference into the licence.
What this means for you
If you hold records in the EU, the UK or Switzerland, work in this order. Sort: separate company-only material from records about people, and leave out what you hold for customers. Shrink: remove sensitive items, staff content and attachments, then de-identify and test. Document the assessments. Tell people and give them time to object. Choose the buyer with the transfer in mind: one that takes a de-identified lot, is listed for your country and will name AI model training as the only purpose. Keep the file, because the buyer's regulator will ask for it.
Buyers do take holders from Europe. Of the 24 buyers of company records whose pages we read, 5 say they take holders from anywhere or work with holders abroad and 15 say nothing about where a holder can be based; which buyers take holders outside the US has the table. European publishers already license to US labs, among them Le Monde, Schibsted and the Guardian with OpenAI and Süddeutsche Zeitung with Meta; of 78 disclosed deals whose licensor's head office we could place, 32 have a licensor outside the US. For tax, see US withholding tax on data licence payments; for rules outside Europe, privacy law by country.
Each country has a page with its rules, tax treaty and buyers: Germany, France, the Netherlands, Ireland, Spain, Italy, Sweden, Poland, Estonia, the United Kingdom and Switzerland, all on the country register. The de-identification tool works on your own computer, and the readiness tool walks through the rights questions. To see what your records might fetch, start with a valuation and read how much company data is worth.
AIData.Broker acts as the holder's agent: the holder licenses its own records to the buyer, and we do not take ownership of them. We run sales only for holders based in the United States, the United Kingdom and Canada today; a valuation works for a holder based anywhere and binds no one. For the rest of the rights picture, see the rights hub, is it legal to sell customer data? and the selling guide.